man, I hate saying I don’t have the brain power right now to find it and I could have sworn I commented on a post discussing this very thing for one of those papers. I know there was recently a paper that contradicts my statement, but a slightly older paper supporting it. I think it is likely a mixed bag at the moment depending on the model and their training regime.
My anecdata of one using anthropic’s and google’s models (google’s especially) this year, the model will drop the casual tone and sycophancy of its replies pretty damn quick as soon as you tell them off. And usually that is when there is less correction. Could also be because my prompting changes from supervisory to very directive, as in shut-up-and-do-exactly-what-I-say, when it gets off into weeds. Even then, it can be a crap shoot.
more like an off-kilter broken human at the moment. currently suffering burnout from being neck deep in this shit from the beginning of the year once the models had that step change into actual utility. Everything from championing the use at work to deliver more in a quarter than the team would have in two, to finishing half a dozen personal projects and starting half a dozen more. Eventually, just looking at the claude code tui would fill me anxiety and every time the model would fuck up, unmitigated rage all while caught in a cycle of drugs that stopped being fun and started becoming dependency. i just couldnt do it any more. it hurts.
man, I hate saying I don’t have the brain power right now to find it and I could have sworn I commented on a post discussing this very thing for one of those papers. I know there was recently a paper that contradicts my statement, but a slightly older paper supporting it. I think it is likely a mixed bag at the moment depending on the model and their training regime.
My anecdata of one using anthropic’s and google’s models (google’s especially) this year, the model will drop the casual tone and sycophancy of its replies pretty damn quick as soon as you tell them off. And usually that is when there is less correction. Could also be because my prompting changes from supervisory to very directive, as in shut-up-and-do-exactly-what-I-say, when it gets off into weeds. Even then, it can be a crap shoot.
Ha, I like that…
lol, yeah, picked that up some data sciencey friends at one point
As a rando that surfed its way towards this convo:
Hell yeah. Either you are a passably off-kilter ai variant, or an unfiltered passionate human. Perhaps an extraterrestrial.
I have mental space for this.
more like an off-kilter broken human at the moment. currently suffering burnout from being neck deep in this shit from the beginning of the year once the models had that step change into actual utility. Everything from championing the use at work to deliver more in a quarter than the team would have in two, to finishing half a dozen personal projects and starting half a dozen more. Eventually, just looking at the claude code tui would fill me anxiety and every time the model would fuck up, unmitigated rage all while caught in a cycle of drugs that stopped being fun and started becoming dependency. i just couldnt do it any more. it hurts.